Apache
/
hadoop
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

HADOOP-7446. Implement CRC32C native code using SSE4.2 instructions. Contributed by Kihwal Lee and Todd Lipcon. 

git-svn-id: https://svn.apache.org/repos/asf/hadoop/common/trunk@1190650 13f79535-47bb-0310-9956-ffa450edef68
This commit is contained in:
Todd Lipcon 2011-10-28 22:51:47 +00:00
parent 6b0963c53b
commit 5839aa01c8
3 changed files with 508 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
hadoop-common-project/hadoop-common/CHANGES.txt
View File

@ -520,6 +520,9 @@ Release 0.23.0 - Unreleased
HADOOP-7753. Support fadvise and sync_file_range in NativeIO. Add
ReadaheadPool infrastructure for use in HDFS and MR. (todd)
HADOOP-7446. Implement CRC32C native code using SSE4.2 instructions.
(Kihwal Lee and todd via todd)
BUG FIXES
HADOOP-7740. Fixed security audit logger configuration. (Arpit Gupta via Eric Yang)

5
hadoop-common-project/hadoop-common/src/main/native/src/org/apache/hadoop/util/NativeCrc32.c
View File

@ -124,6 +124,11 @@ JNIEXPORT void JNICALL Java_org_apache_hadoop_util_NativeCrc32_nativeVerifyChunk
"bad offsets or lengths");
return;
}
if (unlikely(bytes_per_checksum) <= 0) {
THROW(env, "java/lang/IllegalArgumentException",
"invalid bytes_per_checksum");
return;
}
uint32_t *sums = (uint32_t *)(sums_addr + sums_offset);
uint8_t *data = data_addr + data_offset;

506
hadoop-common-project/hadoop-common/src/main/native/src/org/apache/hadoop/util/bulk_crc32.c
View File

@ -21,6 +21,7 @@
* All rights reserved. Use of this source code is governed by a
* BSD-style license that can be found in the LICENSE file.
*/
#include <assert.h>
#include <arpa/inet.h>
#include <stdint.h>
#include <unistd.h>
@ -30,47 +31,124 @@
#include "bulk_crc32.h"
#include "gcc_optimizations.h"
#define USE_PIPELINED
typedef uint32_t (*crc_update_func_t)(uint32_t, const uint8_t *, size_t);
static uint32_t crc_init();
static uint32_t crc_val(uint32_t crc);
static uint32_t crc32_zlib_sb8(uint32_t crc, const uint8_t *buf, size_t length);
static uint32_t crc32c_sb8(uint32_t crc, const uint8_t *buf, size_t length);
#ifdef USE_PIPELINED
static void pipelined_crc32c(uint32_t *crc1, uint32_t *crc2, uint32_t *crc3, const uint8_t *p_buf, size_t block_size, int num_blocks);
#endif USE_PIPELINED
static int cached_cpu_supports_crc32; // initialized by constructor below
static uint32_t crc32c_hardware(uint32_t crc, const uint8_t* data, size_t length);
int bulk_verify_crc(const uint8_t *data, size_t data_len,
const uint32_t *sums, int checksum_type,
int bytes_per_checksum,
crc32_error_t *error_info) {
#ifdef USE_PIPELINED
uint32_t crc1, crc2, crc3;
int n_blocks = data_len / bytes_per_checksum;
int remainder = data_len % bytes_per_checksum;
int do_pipelined = 0;
#endif
uint32_t crc;
crc_update_func_t crc_update_func;
switch (checksum_type) {
case CRC32_ZLIB_POLYNOMIAL:
crc_update_func = crc32_zlib_sb8;
break;
case CRC32C_POLYNOMIAL:
if (likely(cached_cpu_supports_crc32)) {
crc_update_func = crc32c_hardware;
#ifdef USE_PIPELINED
do_pipelined = 1;
#endif
} else {
crc_update_func = crc32c_sb8;
}
break;
default:
return INVALID_CHECKSUM_TYPE;
}
#ifdef USE_PIPELINED
if (do_pipelined) {
/* Process three blocks at a time */
while (likely(n_blocks >= 3)) {
crc1 = crc2 = crc3 = crc_init();
pipelined_crc32c(&crc1, &crc2, &crc3, data, bytes_per_checksum, 3);
crc = ntohl(crc_val(crc1));
if ((crc = ntohl(crc_val(crc1))) != *sums)
goto return_crc_error;
sums++;
data += bytes_per_checksum;
if ((crc = ntohl(crc_val(crc2))) != *sums)
goto return_crc_error;
sums++;
data += bytes_per_checksum;
if ((crc = ntohl(crc_val(crc3))) != *sums)
goto return_crc_error;
sums++;
data += bytes_per_checksum;
n_blocks -= 3;
}
/* One or two blocks */
if (n_blocks) {
crc1 = crc2 = crc_init();
pipelined_crc32c(&crc1, &crc2, &crc3, data, bytes_per_checksum, n_blocks);
if ((crc = ntohl(crc_val(crc1))) != *sums)
goto return_crc_error;
data += bytes_per_checksum;
sums++;
if (n_blocks == 2) {
if ((crc = ntohl(crc_val(crc2))) != *sums)
goto return_crc_error;
sums++;
data += bytes_per_checksum;
}
}
/* For something smaller than a block */
if (remainder) {
crc1 = crc_init();
pipelined_crc32c(&crc1, &crc2, &crc3, data, remainder, 1);
if ((crc = ntohl(crc_val(crc1))) != *sums)
goto return_crc_error;
}
return CHECKSUMS_VALID;
}
#endif
while (likely(data_len > 0)) {
int len = likely(data_len >= bytes_per_checksum) ? bytes_per_checksum : data_len;
uint32_t crc = crc_init();
crc = crc_init();
crc = crc_update_func(crc, data, len);
crc = ntohl(crc_val(crc));
if (unlikely(crc != *sums)) {
if (error_info != NULL) {
error_info->got_crc = crc;
error_info->expected_crc = *sums;
error_info->bad_data = data;
}
return INVALID_CHECKSUM_DETECTED;
goto return_crc_error;
}
data += len;
data_len -= len;
sums++;
}
return CHECKSUMS_VALID;
return_crc_error:
if (error_info != NULL) {
error_info->got_crc = crc;
error_info->expected_crc = *sums;
error_info->bad_data = data;
}
return INVALID_CHECKSUM_DETECTED;
}
@ -154,3 +232,417 @@ static uint32_t crc32_zlib_sb8(
}
return crc;
}
///////////////////////////////////////////////////////////////////////////
// Begin code for SSE4.2 specific hardware support of CRC32C
///////////////////////////////////////////////////////////////////////////
#if (defined(__amd64__) || defined(__i386)) && defined(__GNUC__)
# define SSE42_FEATURE_BIT (1 << 20)
# define CPUID_FEATURES 1
/**
* Call the cpuid instruction to determine CPU feature flags.
*/
static uint32_t cpuid(uint32_t eax_in) {
uint32_t eax, ebx, ecx, edx;
# if defined(__PIC__) && !defined(__LP64__)
// 32-bit PIC code uses the ebx register for the base offset --
// have to save and restore it on the stack
asm("pushl %%ebx\n\t"
"cpuid\n\t"
"movl %%ebx, %[ebx]\n\t"
"popl %%ebx" : "=a" (eax), [ebx] "=r"(ebx), "=c"(ecx), "=d"(edx) : "a" (eax_in)
: "cc");
# else
asm("cpuid" : "=a" (eax), "=b"(ebx), "=c"(ecx), "=d"(edx) : "a"(eax_in)
: "cc");
# endif
return ecx;
}
/**
* On library load, initiailize the cached value above for
* whether the cpu supports SSE4.2's crc32 instruction.
*/
void __attribute__ ((constructor)) init_cpu_support_flag(void) {
uint32_t ecx = cpuid(CPUID_FEATURES);
cached_cpu_supports_crc32 = ecx & SSE42_FEATURE_BIT;
}
//
// Definitions of the SSE4.2 crc32 operations. Using these instead of
// the GCC __builtin_* intrinsics allows this code to compile without
// -msse4.2, since we do dynamic CPU detection at runtime.
//
# ifdef __LP64__
inline uint64_t _mm_crc32_u64(uint64_t crc, uint64_t value) {
asm("crc32q %[value], %[crc]\n" : [crc] "+r" (crc) : [value] "rm" (value));
return crc;
}
# endif
inline uint32_t _mm_crc32_u32(uint32_t crc, uint32_t value) {
asm("crc32l %[value], %[crc]\n" : [crc] "+r" (crc) : [value] "rm" (value));
return crc;
}
inline uint32_t _mm_crc32_u16(uint32_t crc, uint16_t value) {
asm("crc32w %[value], %[crc]\n" : [crc] "+r" (crc) : [value] "rm" (value));
return crc;
}
inline uint32_t _mm_crc32_u8(uint32_t crc, uint8_t value) {
asm("crc32b %[value], %[crc]\n" : [crc] "+r" (crc) : [value] "rm" (value));
return crc;
}
# ifdef __LP64__
/**
* Hardware-accelerated CRC32C calculation using the 64-bit instructions.
*/
static uint32_t crc32c_hardware(uint32_t crc, const uint8_t* p_buf, size_t length) {
// start directly at p_buf, even if it's an unaligned address. According
// to the original author of this code, doing a small run of single bytes
// to word-align the 64-bit instructions doesn't seem to help, but
// we haven't reconfirmed those benchmarks ourselves.
uint64_t crc64bit = crc;
size_t i;
for (i = 0; i < length / sizeof(uint64_t); i++) {
crc64bit = _mm_crc32_u64(crc64bit, *(uint64_t*) p_buf);
p_buf += sizeof(uint64_t);
}
// This ugly switch is slightly faster for short strings than the straightforward loop
uint32_t crc32bit = (uint32_t) crc64bit;
length &= sizeof(uint64_t) - 1;
switch (length) {
case 7:
crc32bit = _mm_crc32_u8(crc32bit, *p_buf++);
case 6:
crc32bit = _mm_crc32_u16(crc32bit, *(uint16_t*) p_buf);
p_buf += 2;
// case 5 is below: 4 + 1
case 4:
crc32bit = _mm_crc32_u32(crc32bit, *(uint32_t*) p_buf);
break;
case 3:
crc32bit = _mm_crc32_u8(crc32bit, *p_buf++);
case 2:
crc32bit = _mm_crc32_u16(crc32bit, *(uint16_t*) p_buf);
break;
case 5:
crc32bit = _mm_crc32_u32(crc32bit, *(uint32_t*) p_buf);
p_buf += 4;
case 1:
crc32bit = _mm_crc32_u8(crc32bit, *p_buf);
break;
case 0:
break;
default:
// This should never happen; enable in debug code
assert(0 && "ended up with 8 or more bytes at tail of calculation");
}
return crc32bit;
}
#ifdef USE_PIPELINED
/**
* Pipelined version of hardware-accelerated CRC32C calculation using
* the 64 bit crc32q instruction.
* One crc32c instruction takes three cycles, but two more with no data
* dependency can be in the pipeline to achieve something close to single
* instruction/cycle. Here we feed three blocks in RR.
*
* crc1, crc2, crc3 : Store initial checksum for each block before
* calling. When it returns, updated checksums are stored.
* p_buf : The base address of the data buffer. The buffer should be
* at least as big as block_size * num_blocks.
* block_size : The size of each block in bytes.
* num_blocks : The number of blocks to work on. Min = 1, Max = 3
*/
static void pipelined_crc32c(uint32_t *crc1, uint32_t *crc2, uint32_t *crc3, const uint8_t *p_buf, size_t block_size, int num_blocks) {
uint64_t c1 = *crc1;
uint64_t c2 = *crc2;
uint64_t c3 = *crc3;
uint64_t *data = (uint64_t*)p_buf;
int counter = block_size / sizeof(uint64_t);
int remainder = block_size % sizeof(uint64_t);
uint8_t *bdata;
/* We do switch here because the loop has to be tight in order
* to fill the pipeline. Any other statement inside the loop
* or inbetween crc32 instruction can slow things down. Calling
* individual crc32 instructions three times from C also causes
* gcc to insert other instructions inbetween.
*
* Do not rearrange the following code unless you have verified
* the generated machine code is as efficient as before.
*/
switch (num_blocks) {
case 3:
/* Do three blocks */
while (likely(counter)) {
__asm__ __volatile__(
"crc32q (%7), %0;\n\t"
"crc32q (%7,%6,1), %1;\n\t"
"crc32q (%7,%6,2), %2;\n\t"
: "=r"(c1), "=r"(c2), "=r"(c3)
: "r"(c1), "r"(c2), "r"(c3), "r"(block_size), "r"(data)
);
data++;
counter--;
}
/* Take care of the remainder. They are only up to three bytes,
* so performing byte-level crc32 won't take much time.
*/
bdata = (uint8_t*)data;
while (likely(remainder)) {
__asm__ __volatile__(
"crc32b (%7), %0;\n\t"
"crc32b (%7,%6,1), %1;\n\t"
"crc32b (%7,%6,2), %2;\n\t"
: "=r"(c1), "=r"(c2), "=r"(c3)
: "r"(c1), "r"(c2), "r"(c3), "r"(block_size), "r"(bdata)
);
bdata++;
remainder--;
}
break;
case 2:
/* Do two blocks */
while (likely(counter)) {
__asm__ __volatile__(
"crc32q (%5), %0;\n\t"
"crc32q (%5,%4,1), %1;\n\t"
: "=r"(c1), "=r"(c2)
: "r"(c1), "r"(c2), "r"(block_size), "r"(data)
);
data++;
counter--;
}
bdata = (uint8_t*)data;
while (likely(remainder)) {
__asm__ __volatile__(
"crc32b (%5), %0;\n\t"
"crc32b (%5,%4,1), %1;\n\t"
: "=r"(c1), "=r"(c2)
: "r"(c1), "r"(c2), "r"(c3), "r"(block_size), "r"(bdata)
);
bdata++;
remainder--;
}
break;
case 1:
/* single block */
while (likely(counter)) {
__asm__ __volatile__(
"crc32q (%2), %0;\n\t"
: "=r"(c1)
: "r"(c1), "r"(data)
);
data++;
counter--;
}
bdata = (uint8_t*)data;
while (likely(remainder)) {
__asm__ __volatile__(
"crc32b (%2), %0;\n\t"
: "=r"(c1)
: "r"(c1), "r"(bdata)
);
bdata++;
remainder--;
}
break;
case 0:
return;
default:
assert(0 && "BUG: Invalid number of checksum blocks");
}
*crc1 = c1;
*crc2 = c2;
*crc3 = c3;
return;
}
#endif /* USE_PIPELINED */
# else // 32-bit
/**
* Hardware-accelerated CRC32C calculation using the 32-bit instructions.
*/
static uint32_t crc32c_hardware(uint32_t crc, const uint8_t* p_buf, size_t length) {
// start directly at p_buf, even if it's an unaligned address. According
// to the original author of this code, doing a small run of single bytes
// to word-align the 64-bit instructions doesn't seem to help, but
// we haven't reconfirmed those benchmarks ourselves.
size_t i;
for (i = 0; i < length / sizeof(uint32_t); i++) {
crc = _mm_crc32_u32(crc, *(uint32_t*) p_buf);
p_buf += sizeof(uint32_t);
}
// This ugly switch is slightly faster for short strings than the straightforward loop
length &= sizeof(uint32_t) - 1;
switch (length) {
case 3:
crc = _mm_crc32_u8(crc, *p_buf++);
case 2:
crc = _mm_crc32_u16(crc, *(uint16_t*) p_buf);
break;
case 1:
crc = _mm_crc32_u8(crc, *p_buf);
break;
case 0:
break;
default:
// This should never happen; enable in debug code
assert(0 && "ended up with 4 or more bytes at tail of calculation");
}
return crc;
}
#ifdef USE_PIPELINED
/**
* Pipelined version of hardware-accelerated CRC32C calculation using
* the 32 bit crc32l instruction.
* One crc32c instruction takes three cycles, but two more with no data
* dependency can be in the pipeline to achieve something close to single
* instruction/cycle. Here we feed three blocks in RR.
*
* crc1, crc2, crc3 : Store initial checksum for each block before
* calling. When it returns, updated checksums are stored.
* data : The base address of the data buffer. The buffer should be
* at least as big as block_size * num_blocks.
* block_size : The size of each block in bytes.
* num_blocks : The number of blocks to work on. Min = 1, Max = 3
*/
static void pipelined_crc32c(uint32_t *crc1, uint32_t *crc2, uint32_t *crc3, const uint8_t *p_buf, size_t block_size, int num_blocks) {
uint32_t c1 = *crc1;
uint32_t c2 = *crc2;
uint32_t c3 = *crc3;
int counter = block_size / sizeof(uint32_t);
int remainder = block_size % sizeof(uint32_t);
uint32_t *data = (uint32_t*)p_buf;
uint8_t *bdata;
/* We do switch here because the loop has to be tight in order
* to fill the pipeline. Any other statement inside the loop
* or inbetween crc32 instruction can slow things down. Calling
* individual crc32 instructions three times from C also causes
* gcc to insert other instructions inbetween.
*
* Do not rearrange the following code unless you have verified
* the generated machine code is as efficient as before.
*/
switch (num_blocks) {
case 3:
/* Do three blocks */
while (likely(counter)) {
__asm__ __volatile__(
"crc32l (%7), %0;\n\t"
"crc32l (%7,%6,1), %1;\n\t"
"crc32l (%7,%6,2), %2;\n\t"
: "=r"(c1), "=r"(c2), "=r"(c3)
: "r"(c1), "r"(c2), "r"(c3), "r"(block_size), "r"(data)
);
data++;
counter--;
}
/* Take care of the remainder. They are only up to three bytes,
* so performing byte-level crc32 won't take much time.
*/
bdata = (uint8_t*)data;
while (likely(remainder)) {
__asm__ __volatile__(
"crc32b (%7), %0;\n\t"
"crc32b (%7,%6,1), %1;\n\t"
"crc32b (%7,%6,2), %2;\n\t"
: "=r"(c1), "=r"(c2), "=r"(c3)
: "r"(c1), "r"(c2), "r"(c3), "r"(block_size), "r"(bdata)
);
bdata++;
remainder--;
}
break;
case 2:
/* Do two blocks */
while (likely(counter)) {
__asm__ __volatile__(
"crc32l (%5), %0;\n\t"
"crc32l (%5,%4,1), %1;\n\t"
: "=r"(c1), "=r"(c2)
: "r"(c1), "r"(c2), "r"(block_size), "r"(data)
);
data++;
counter--;
}
bdata = (uint8_t*)data;
while (likely(remainder)) {
__asm__ __volatile__(
"crc32b (%5), %0;\n\t"
"crc32b (%5,%4,1), %1;\n\t"
: "=r"(c1), "=r"(c2)
: "r"(c1), "r"(c2), "r"(c3), "r"(block_size), "r"(bdata)
);
bdata++;
remainder--;
}
break;
case 1:
/* single block */
while (likely(counter)) {
__asm__ __volatile__(
"crc32l (%2), %0;\n\t"
: "=r"(c1)
: "r"(c1), "r"(data)
);
data++;
counter--;
}
bdata = (uint8_t*)data;
while (likely(remainder)) {
__asm__ __volatile__(
"crc32b (%2), %0;\n\t"
: "=r"(c1)
: "r"(c1), "r"(bdata)
);
bdata++;
remainder--;
}
break;
case 0:
return;
default:
assert(0 && "BUG: Invalid number of checksum blocks");
}
*crc1 = c1;
*crc2 = c2;
*crc3 = c3;
return;
}
#endif /* USE_PIPELINED */
# endif // 64-bit vs 32-bit
#else // end x86 architecture
static uint32_t crc32c_hardware(uint32_t crc, const uint8_t* data, size_t length) {
// never called!
assert(0 && "hardware crc called on an unsupported platform");
return 0;
}
#endif